While the performance impact of these changes are in the noise under
macOS 10.12.3 on a on a MacBook Pro with a 2.9GHz Intel Core i5 CPU and
8GB 1867MHz DDR3 RAM, eliminating these subshells makes the code more
consistent.
I did try removing `buffer` to eliminate yet more subshells, but the
flickering of the output did prove annoying, so I'm not removing it.
This also update the behaviour of the summary, now it only display the
number of failures, and skipped tests also, if the numbers are greater
than zero.
The `((x++))` syntax is shorthand for `let x++`. According to `help let`:
If the last ARG evaluates to 0, let returns 1; 0 is returned
otherwise.
Thus the exit status of the expression `x=0; let x++` is 1, since the post-increment `++` operator evaluates to the value of the variable before incrementing.
In Bash 4, this non-zero exit status properly triggers `set -e`'s error trap, but in Bash 3 it does not. That's why the tests were passing on OS X (Bash 3) but not Linux (Bash 4).
We can work around the problem by choosing an incrementation expression that never evaluates to 0, such as `+=` or the pre-increment `++` operator. For consistency and clarity, I've changed to `x+=1` everywhere.
Ref. #25, #27
